DeepSeek-V4

A new generation of open source large models released by Chinese AI companies, with 1M ultra-long context and agent capabilities comparable to top closed source models.

large language model

https://api-docs.deepseek.com/news/news260424

Visit Website

In-depth Report

DeepSeek-V4 is a preview version of a new generation of large language models released by the Chinese artificial intelligence company DeepSeek on April 24, 2026, with simultaneous open source model weights and technical reports. The product includes two versions, Pro and Flash. The entire series comes standard with 1M (million tokens) ultra-long context. It has reached the world's top open source model level in terms of agent capabilities, code generation and complex reasoning, and its pricing is only about one-seventh of that of overseas closed-source models of the same level. The release of V4 marks that China's AI open source power occupies an important position in the global large-scale model competition. At the same time, it has special industrial significance because of its explicit support for domestic Huawei Ascend chips.
Founded in 2023, DeepSeek is a Chinese startup company focusing on the research of underlying models and technologies of general artificial intelligence. The company is headquartered in China and its founder is Liang Wenfeng. According to public information, DeepSeek shocked the U.S. AI industry with the release of the R1 model in early 2025 and is regarded as a representative company for China's AI technology breakthroughs. The company conducts model training based on its self-developed training framework, self-built intelligent computing clusters and Wanka computing resources. On April 24, 2026, DeepSeek released the highly anticipated V4 model about three years after its establishment, which is exactly one year since the last R1 release. It is worth noting that DeepSeek is currently actively deploying physical infrastructure and launching large-scale recruitment for intelligent computing centers in Ulanqab, Inner Mongolia, demonstrating its strategic intention to extend upstream of the industry chain. In terms of financing, although the specific amount has not been publicly disclosed, industry rumors say that DeepSeek's target valuation has reached US$20 billion, which reflects the capital market's high recognition of its technical strength and business prospects. The company has always adhered to the open source route during its development. The previous V3 and R1 models have been widely used in the global developer community. According to OpenRouter's public statistics, the token throughput of these two models once accounted for more than half of all open source models.
DeepSeek-V4 has achieved many breakthroughs in technical architecture and functional features. The first is the universalization of ultra-long context capabilities. The entire product series comes standard with 1M (million tokens) ultra-long context, which is an 8-fold improvement compared to the 128K of V3.2. This means that users can directly import complete code libraries, ultra-long industry documents, and complete books with millions of words for end-to-end processing, which greatly simplifies the technical link of long text processing. Users do not need to build an additional RAG system to achieve efficient information retrieval and understanding. In terms of model architecture, V4-Pro uses a MoE (Mixed Expert) architecture with a total parameter of 1.6T and an activation parameter of 49B, and the amount of pre-training data reaches 33T. V4-Flash adopts a lightweight design with an activation parameter of 13B, and achieves extremely low latency and high cost performance through a new token compression attention mechanism and DSA sparse attention architecture. According to the test data officially released by DeepSeek, V4-Pro surpassed all publicly evaluated open source models in core reasoning evaluations such as mathematics, STEM, and competition-level code, and reached the best level of open source models in the Agentic Coding evaluation. The delivery quality is close to Claude Opus 4.6 non-thinking mode. Internal testing shows that the usage feedback of V4-Pro is better than Anthropic Sonnet 4.5, and it has become the main agent coding tool for DeepSeek internal employees. V4 also supports dual-mode switching between non-thinking mode and thinking mode. Users can customize the thinking intensity through the reasoning_effort parameter. The product fully supports Json Output, Tool Calls, and conversation prefix continuation capabilities, and has been specifically adapted and optimized for mainstream Agent products such as Claude Code, OpenClaw, OpenCode, and CodeBuddy. In terms of open source and deployment, model weights are simultaneously open for download on the Hugging Face and ModelSpace platforms, fully retaining all capabilities consistent with the official cloud API without any functional castration. It also supports the full process tool chain of model fine-tuning, quantification, and inference acceleration, and has completed Day 0 native adaptation of mainstream inference frameworks such as vLLM and TGI, and Agent frameworks such as LangChain and LlamaIndex.In terms of access method, users do not need to modify the original base_url, and only need to replace the model parameter with the corresponding version name. V4 is compatible with both OpenAI ChatCompletions and Anthropic interface formats, significantly reducing migration costs. It is worth noting that the old API interfaces (deepseek-chat, deepseek-reasoner) will be discontinued on July 24, 2026.
DeepSeek-V4 adopts a tiered pricing strategy and provides two versions, Pro and Flash, to meet the needs of different user groups. The pricing of the V4-Flash version is extremely user-friendly: the input price for cache hits is 0.2 yuan/million tokens, the input price for cache misses is 1 yuan/million tokens, and the output price is 2 yuan/million tokens. The pricing of the V4-Pro version is positioned at the mid-to-high end: the input price for cache hits is 1 yuan/million tokens, the input price for cache misses is 12 yuan/million tokens, and the output price is 24 yuan/million tokens. Compared with overseas closed-source models of the same level, V4’s pricing has significant advantages. Taking GPT-5.5 as an example, its input price is about US$5/million tokens and its output price is about US$30/million tokens, while the corresponding prices of V4-Pro are only 12 yuan and 24 yuan (about 1.7 US dollars and 3.4 US dollars), which is about one-third to one-tenth of GPT-5.5. This pricing strategy continues DeepSeek's consistent "high cost performance" line, aiming to turn top large model capabilities from a tool for a few people into an infrastructure that more people can call upon. From a business model perspective, DeepSeek adopts a dual-track strategy of open source free + cloud payment. Model weights are open source and free meaning that anyone can download and deploy them locally, which is very attractive to enterprises and developers. Cloud API services obtain revenue through pay-as-you-go billing. This model not only meets the needs of business users who need quick access, but also establishes a broad developer ecosystem through open source.
Judging from community feedback, the release of DeepSeek-V4 has attracted great attention from the developer community. On the first day of release, Zhihu's single-day popularity reached 53.9 million, Hacker News received 999 likes, Reddit received more than 600 likes, and multiple platforms simultaneously detonated. In the technical community, developers generally have positive evaluations of V4, especially its performance in long context processing, code generation and cost-effectiveness. Positive comments are mainly concentrated in the following aspects: first, the ultra-long context capability is widely recognized, and the 1M context standard allows developers to process large-scale text tasks without additional costs; second, the price/performance ratio is extremely competitive, and the price of the Flash version is almost negligible for small and medium-sized projects; third, the open source strategy is well received, and the practice of fully retaining all functions without any castration reflects DeepSeek's sincerity to the open source community; fourth, the API compatibility is excellent, and it is compatible with OpenAI and Anthropic interface formats, which greatly reduces migration costs. However, there are also some skeptical voices. Some users are concerned that V4 took a long time from announcement to official release (called "delay" in the market), and believe that this may reflect the challenges the team faces in technology research and development. In addition, as a newly released product, the stability and long-term performance of V4 in a real production environment still need to be verified.
From an industry perspective, the release of DeepSeek-V4 has multiple meanings. First of all, at the technical level, V4 has proven that open source models can be close to or even comparable to top closed source models in performance, which has had an important impact on the technological development path of the entire AI industry. Replit CEO Amjad Masad and other American developers spoke highly of V4's attention compression and long context efficiency improvements, believing that this represents the highest level of the current open source model. Secondly, at the industrial level, V4 clearly supports domestic Huawei Ascend chips, which is of great significance to China’s local AI chip ecosystem. MIT Technology Review China Edition reported that DeepSeek is building a “unified inference middleware” with the goal of reconstructing the domestic large model technology stack. If the performance of V4 on Ascend chips is close to that of NVIDIA GPUs, it will be a major breakthrough for the domestic AI chip ecosystem. Third, in terms of market competition, the release of V4 has intensified the competition in the field of large models. It is worth noting that the day V4 was released coincided with the release of GPT-5.5 by OpenAI, and the flagship products of the two companies competed head-on. This is not only a technical showdown between China’s open source power and the peak of American closed source, but also indicates that the large model market will become more diversified and competitive in the future. In terms of international media, authoritative technology media such as The Verge have given positive reports on V4. They believe that DeepSeek's resurgence one year after the release of R1 indicates that Chinese AI companies have the strength to compete head-on with American competitors on the global stage.
Although DeepSeek-V4 has received widespread market attention and positive reviews, it also faces some controversy and potential risks. In terms of technical disputes, the multiple delays in the release of V4 have caused the market to question its research and development progress. From the originally expected October 2025 to the official release in April 2026, the interval is as long as half a year, which affects market expectations to a certain extent. Some analysts believe that the delayed release may be related to the team’s ongoing architectural upgrade, but there are also concerns about whether this means that the technology has encountered a bottleneck. In terms of geopolitics, DeepSeek has previously attracted attention from the United States. U.S. officials have reportedly accused DeepSeek of using banned Nvidia chips, and Anthropic has also claimed that DeepSeek misused Claude to improve its own products. Although these accusations have not been confirmed, they may cause potential obstacles to DeepSeek's overseas market expansion. In terms of technical risks, as a newly released product, the stability and reliability of V4 in a large-scale production environment still need to be verified. In addition, as the user scale grows, how to maintain service quality and service response speed will be a practical challenge faced by the team. In terms of business models, although the open source strategy helps build a developer ecosystem, it also means that direct revenue sources are relatively limited. How to find a balance between open source promotion and commercial monetization will be a question that DeepSeek needs to continue to think about.
Based on the product features of DeepSeek-V4, the following groups of people may gain the greatest value from it. For AI developers and researchers, V4’s open source nature, complete tool chain support, and compatibility with mainstream frameworks make it an ideal research platform. The 1M ultra-long context capability is particularly suitable for developers who need to perform code analysis, document processing, academic research, etc. For start-ups and small and medium-sized companies, the V4-Flash version provides a very cost-effective option. Its extremely low API call cost and excellent performance can help teams with limited resources quickly access advanced large model capabilities without bearing high infrastructure costs. For enterprise users who need to process long documents, V4's standard 1M context eliminates dependence on the RAG system and can directly process large-scale texts such as complete books, reports, code libraries, etc., simplifying the technical architecture and reducing maintenance costs. For enterprises that are concerned about data privacy and security, the open source feature of V4 allows for completely private deployment without data leaving the enterprise's own servers. This provides a feasible solution for industries that have strict requirements for data security. In terms of usage suggestions, you need to pay attention to the following points: first, migrate the old API interface in a timely manner, because the old interface will be deactivated on July 24, 2026; second, choose the appropriate version according to actual needs. For lightweight high-frequency scenarios, you can choose the Flash version, and for scenarios that require stronger reasoning capabilities, it is recommended to choose the Pro version; third, pay close attention to official updates, V4 as a new product may continue to launch optimized versions.
The release of DeepSeek-V4 is an important milestone for China’s AI open source power in the global large model competition. With 1M ultra-long context, world-class agent capabilities and extremely competitive pricing, V4 has successfully changed the top large model capabilities from "optional" to "standard", promoting the inclusive process of AI technology. Its strategic choice of clearly supporting domestic chips not only reflects its technological confidence, but also injects confidence into the development of the domestic AI chip ecosystem. In the long run, the successful release of V4 indicates that Chinese AI companies are transforming from technology followers to innovation leaders. In the context of increasingly fierce global AI competition, the open source route adhered to by DeepSeek may have a profound impact on the development pattern of global AI technology. Of course, as a new product, V4 still needs to withstand the test of the market and time. Its performance in the production environment, the development of subsequent versions, and the verification of the business model all deserve continued attention.

User Reviews

q88n2d

—

Grayscale has tested the official version of V4 for a few days. The overall version is close to the Opus 4.8 level. The encoding is catching up with GPT-5.6 Sol. The agent capabilities are a qualitative leap compared to the previous generation. The improvement in 3D scene and SVG rendering is visible to the naked eye, but the number of iteration rounds for the same task is still more than that of Fable 5. Considering the price, I really can't fault it.

MIvai

—

The grayscale test report is out. The official version of V4’s thinking chain shows stronger characteristics of Agent’s independent planning. It does not just write code, but advances independently in the complete cycle of planning, construction and debugging. Looking forward to full delivery.

CarlBaker

—

Using V4 Pro to run the Electron build, the same configuration item was changed repeatedly and jumped repeatedly. GLM 5.1 locates the root cause in one round and corrects it without recurrence. The depth gap in engineering context tracking is still real.

ZacharyMendoza_2023

—

V4 million context is standard on all series, but don’t rush to Pro. The Flash version is sufficient for daily use, and the sudden endless loop problem of V3 is basically gone on Flash.

Patrick_Watson_9973

—

The front-end aesthetic of the preview version is really awesome, with the whole screen in purple color. However, the official version of SVG and 3D has been greatly improved. Xbox controllers and CSGO games can be generated and run at once. This is the level it should be.

Christian.MendozaSr

—

The official version of SWE-bench is only 0.2 points away from Opus 4.6 Max. The stability of the same prompt is good and bad, but there is still a gap with Opus. But the price is one-seventh of Opus, so what kind of bike do you want?

6za_1t

—

The 3D scene production is much better than the preview version. The famous 3D scene that flipped over in the preview version will not appear again in the official version. The Agent completion rate has increased by 21%, and multi-modal and SVG rendering capabilities have been fully supplemented. From the architecture verification of the preview version to the burst of capabilities of the official version, the wait for three months was not in vain.

Raymond_Richardson_20204

—

The official version is in full volume! Native multimodal reasoning is no longer pure text. Although it is not as good as the visual understanding of GPT-5.5, OCR plus image reasoning is enough for daily use. From architecture verification to capability explosion, the wait was not in vain for three months.

الناکوتی

—

V4 crashed again for a day and the server was busy. It was probably a stress test of the official version. Peak and valley pricing is coming. The cost of running the Agent during the day will double. In the future, batch tasks will have to be moved to nighttime. The cache hit price is extremely low, and the peak staggering cost of batch tasks is affordable.

Jacob249

—

I received an email saying that the two old interfaces deepseek-chat and deepseek-reasoner will be deactivated on July 24. The official version also has peak and valley pricing doubled. Cheap products have also begun to install meter, which is a bit sad.

EFisher_2022

—

I don’t agree with NIST’s statement that V4 is 8 months behind the United States. V4 long context efficiency is clearly industry-leading. 1M token inference FLOPs are only 27% of V3.2, and KV cache is reduced to 10%. The official self-evaluation also said that the gap is about 3 to 6 months instead of 8 months. It takes the low-cost and inclusive route instead of extreme computing power, and the evaluation criteria are different.

DMurphy_660

—

V4 Full Paper is out, FP4 QAT post-training is run directly, MoE expert weights are quantized to FP4, and inference is run directly with FP4 weights. QK selector accelerates 2 times and 99.7% recall retention. Being able to achieve this compression ratio with minimal quality degradation would significantly change the cost structure.

clPER

—

Use V4 to write quantitative strategies. The code is neat and orderly, and you will lose money all the way. The analysis on paper makes sense, but once exposed to the noise and fluctuations of the real market, it becomes timid. The model understands the strategy but not the market.

Aaron_Lopez369

—

10 practitioners chatted for a while, and the conclusion was quite consistent: a one-shot success rate of several thousand lines of code is high, and it is sufficient for small applications. However, tool call stability and hallucination rate are flaws, and long-chain tasks cannot be matched with the Harness framework. V4 is an accelerator in the Agent era but it is not a silver bullet. Its implementation requires scaffolding. Flash has superior performance for the same size and will become the first choice base for post-training.

CRivera007

—

The actual test V4 sorted out the relationship between the characters in "The Romance of the Three Kingdoms" and listed the reasons for the deaths of generals and weapons in a table without any overstatement. Financial practitioners can read the annual report through more than 300 pages of PDF and extract detailed data in one go, which greatly increases their productivity. It’s just that the UI aesthetic is really bad.

LisaCarter00787

—

15 projects were tested on V4 Pro, from Lu Xun's imitation to the official SaaS website. The code is indeed stronger. The most surprising thing is that Chinese writing is much more natural than GPT-5.5, without the pretentious artificiality of AI. Expert mode is obviously a notch better.

GeorgeClark

—

The national supercomputer has launched V4 free dialogue, with full-link domestic adaptation. After the Ascend 950 Pro is launched in the second half of the year, the price of the Pro will be significantly reduced. Close-source manufacturers really cannot withstand this price war.

CHrtd

—

It’s true that V4’s coding capabilities are comparable to Opus. You can write a writer recommendation tool for 40 yuan in a second. V4 is free and you can rewrite thousands of lines of code. The cost advantage is exaggerated.

lUCIAbAKER

—

DeepSeek internally switched the main model to V4, and the feedback is better than Sonnet 4.5 and close to Opus 4.6. It is better to eat your own dog food first than many companies.

JJones007

—

Millions of contexts come standard! Find abnormal passages in a lost 1 million word novel and locate them in seconds. Gemini's previously exclusive trump card is now available to you for free, which is truly inclusive.

ViiviNiskanen

—

The bottom layer of V4 has been fully migrated from CUDA to Huawei's CANN framework, and Ascend chips run directly. In the second half of the year, the inference cost of 950 super nodes will be reduced when it is launched. This is the real trump card.

JacquelineJohnson66

—

The Flash version's output of one million tokens is only 0.28 US dollars, which is less than 1% of Claude Opus. Code Arena has defeated the GPT-5.4 high-intensity mode. GPT-5.5 Sorry, V4 is the new shocking moment.

SophiaCruzK497

—

Conway's Game of Life generates runnable seeds at the click of a mouse. It is slower than GPT-5.3 but has much better visual aesthetics. Each has its own merits.

悠然_3

—

The "desperate father"'s genetic red-green color blindness overturned, and V4 did not get the key points in one round. The boundaries of text training for embodied common sense reasoning are still very obvious.

iddpmzb_n9

—

The car wash issue is finally right! V3 allows you to calculate coordinates and recommends pushing a cart there, but V4 Flash can walk there in seconds. This is what normal people would say.

Michelle_Gonzalez

—

It is so important to have zero illusions when citing laws. When asked about the third paragraph of Article 38 of the "Network Data Security Management Regulations", V4 directly said that this article only has two paragraphs and no third paragraph, and also actively guided it to the truly relevant Articles 26, 36 and other clauses. This kind of responsible ignorance is ten thousand times better than making up nonsense.

NazárioAlves

—

The price of the Flash version is really good, 0.2 yuan/million tokens, what kind of bicycle do you want at this price?

JacobJohnson_Pro

—

When trying to find a needle in a haystack and insert a marker into a million-word novel, I failed a few times. After reducing it to 500,000 words, the accuracy rate improved significantly. The official said that 1M but less than 500,000 is the most stable, and it is recommended to split if it exceeds 1M.

SUhar

—

Using V4 Pro with Claude Code to do the project, the total cost for the four tasks in one day was 8 yuan. Comparing the pricing of Claude Opus, this cost gap is outrageous. If the illusion and tool stability can be complemented at the Harness layer, the economic inflection point of Agent will be here.

GaryAlvarez_77

—

The wait is finally here! The 1M context is directly filled up, and you no longer need to bother with RAG.

DeepSeek-V4

In-depth Report

User Reviews

Similar Products